Launch and drive your agents
In an open project, the Agents section of the project bar is where your agents work. It contains tabs, and each tab contains panes, one pane per agent. You launch everything at once, talk to each one separately, and follow their progress in the header of their pane.
Before you start
Section titled “Before you start”- A project open in the app. See Create your first project.
- At least one agent connected. See Connect your AI account.
- Roblox Studio open on the project’s place if you want your agents to work in it. See Link a place to your project.
Add agents and launch
Section titled “Add agents and launch”- In an empty tab, the app shows “Launch an agent on this project,” or “Launch without a preset” if you already have presets. The large Launch name button immediately launches your favorite agent. If you do not have one, it launches the first installed agent, or Claude Code.
- To put together a team, use the Pick another agent menu instead: one click adds a pane without launching anything. Once an agent is added, the Add an agent menu in the dock adds more. Agents pinned to the project bar can also be added with one click.
- Or choose a saved preset, which sets up the team and its layout in one go. See Presets.
- Check the Before launch bar above the grid. It shows the settings that will apply to this launch.
- Click the button in the dock, at the bottom of the grid: Launch N agents (or Open the terminal when the grid contains only terminals). The banner shows “Launching agents…” and “You can keep working.”
- If an agent is not connected, the Connect your agents box opens (never for OpenCode, which is configured in its own terminal). Connect it on the spot: “the launch resumes on its own.” Or click Launch anyway: the agent will then ask you to connect in its terminal.
The launch dock appears only once an agent has been added to the tab, and it disappears as soon as the agents are running. With the large button on the empty tab, it is the only launch button: the project bar no longer has one.
The “Before launch” bar
Section titled “The “Before launch” bar”Above the grid, as long as nothing is running, the Before launch bar shows the choices that will apply. It disappears once the agents are launched: these settings are decided beforehand, not during the run.
| Setting | Options |
|---|---|
| Assets | Search, Hybrid, or Generation |
| 3D models | Auto, Blender, or Remote |
| Preview before adding | Each generated asset is shown to you before it enters the game |
The app recommends the Hybrid mode for assets: “The agents search the bank first, then generate what is missing.” It recommends the Auto mode for 3D models. The All settings link opens the full page. See Review what the agent proposes and 3D models.
Talk to an agent
Section titled “Talk to an agent”Click in the agent’s pane and type your request, in French. It replies in the language you chose in VibeStarter, even if it resumes a conversation started in English. The code itself stays in English.
To type the same thing in every pane at the same time, activate
Broadcast keystrokes.
Follow what an agent is doing
Section titled “Follow what an agent is doing”The header of each pane is kept up to date by the agent itself. It shows:
- a mission title of two to four words;
- an activity line that says what it is doing right now;
- a task list of three to eight steps, each one to do, in progress, or done.
Three states can appear: “Plan” while it is preparing, “Waiting for your reply” when it hands control back to you, and “Plan complete” at the end. An important milestone also triggers a notification.
Pane actions
Section titled “Pane actions”Once the agent is launched, the header of its pane shows these buttons. When the pane is too narrow, they move into the Pane actions menu.
| Action | What it does |
|---|---|
| Mention a project (AI agents, not the Terminal) | Gives the agent another of your projects as a read-only reference |
| Maximize pane / Exit maximize | Makes this pane full screen, and gets out of full screen |
| Broadcast keystrokes | Writes to all panes at once |
| Restart and resume agent | Restarts the agent and has it pick up its conversation again |
| Larger font / Smaller font | Changes the text size in this pane |
The pane’s X, Stop the agent, stops only this agent and removes its pane, without confirmation. Before launch, it is called Remove agent. The X on a tab, Close tab, closes the tab and its panes.
Keyboard shortcuts
Section titled “Keyboard shortcuts”Shortcuts are set in the settings, in the Shortcuts tab, in the
Agent keybindings section. They all work in two steps: the Ctrl + B combination, then an action key.
- Up, Down, Move Left, Move Right: move focus to another pane.
- Toggle Zoom: enlarge or shrink the active pane.
- Toggle Broadcast: “Write to all panes at once.”
- Next Tab, Previous Tab, New tab.
- Toggle Notes: open or close the notes panel.
Two other settings are in the same place: “Text Macros,” which type a prepared text into the active terminal, with an option to send it automatically by adding the Enter key, and the “Command Palette,” available from any screen.
The details are in Keyboard shortcuts.
Drop files into an agent
Section titled “Drop files into an agent”Drag a file or a folder onto an agent’s pane: the area shows “Drop files or a folder
here,” then “Copying into the agent…” At the end, the app tells you how many files
were copied and how many items were skipped. Build folders, project history, and
symbolic links are not copied. A key or credentials file (.env, SSH keys,
certificates) is never sent: inside a folder, it is skipped; dropped or pasted on
its own, it is refused.
How many agents at once
Section titled “How many agents at once”The number of agents depends on your computer’s memory, not on an option you buy. These are the guidelines the app shows at first launch:
| Memory | What the app says |
|---|---|
| Less than 6 GB | Below the minimum |
| 6 to 7 GB | “risky” |
| 8 to 15 GB | “enough, but several agents at once will be tight” |
| 16 GB or more | Recommended |
You also need at least two processor threads and 5 GB of free disk space. Below 20 GB of free space, the app warns that project backups will be tight. The details are in Check that your computer is compatible.
When your agents wait for Studio
Section titled “When your agents wait for Studio”Only one agent drives a Roblox place at a time. This is not a failure: the app shows it plainly with “Studio busy” followed by the name of the agent doing the work, the number of agents waiting, and “Waiting for a place” with the position in the queue.
The turn frees up on its own shortly after the last action of the agent that held the place. To take back control during a test, click the Stop · place name badge in the header of the pane of the agent that is testing: it stops this test and releases its only lock.
Meanwhile, the waiting agents keep working on whatever does not require Studio. See What agents can do in Studio.
When memory runs out
Section titled “When memory runs out”The app sets some agents aside on its own to stay within the available memory. Two badges can then appear, and they are normal:
- “Paused — resumes automatically,” and in the pane “Optimizing resources… resumes automatically”;
- “Queued — starts automatically,” meaning “Waiting for memory before starting.”
A third message does not pause anything. Under an agent’s terminal, “Out of memory (… GB): this agent is slowing down. Restarting it frees memory.” means that this agent alone has filled the share of memory reserved for it. It keeps working, more slowly, and the other agents are not affected. Close its pane, then launch it again: it starts with empty memory. The message disappears on its own if the agent frees its memory.
Do not confuse three compilation messages:
- “Out of memory — The compiler has run out of the virtual machine’s memory. This is not a code error: no agent will fix it. Close some agents to free memory, then reopen the project.”
- “Compile fix in progress — The code doesn’t compile yet: ask an agent to fix the error.” This message appears only if a code agent is running in the project. An agent that is working sees the error and fixes it on its own; an idle agent waits for you to ask it.
- “The code doesn’t compile — The code doesn’t compile: sync with Studio resumes as soon as it does. Launch an agent to fix the error, or fix the file yourself.” No agent is running: it is up to you to launch one.
Limits
Section titled “Limits”- One agent per Roblox place at a time, and only one test at a time per experience.
- Two agents on the same task do not go faster, and they use up your quota twice.
- Closing the app stops all sessions. Pause them instead. See Pause and resume a session.
- An agent tab belongs to a single project. For another project, open it alongside: see Several projects at once.
Frequently asked questions
Section titled “Frequently asked questions”“I can’t type anywhere.” The panes accept text only once the agents are launched. First click the button in the dock, then type in the pane.
“How many agents can I launch?” As many as your memory allows. The only real rule is not to put two agents on the same task.
“Why does ‘Optimizing resources’ keep coming back?” It is the memory safeguard. Reduce the number of agents you launch and close other heavy applications.
“Can the agent work while I am away?” Yes, as long as the app and Roblox Studio stay open and your quota lasts. It stops if it needs an answer from you: its pane then shows “Waiting for your reply.”
“What is a second tab for?” To organize your screen, and to run another team of agents in parallel. Another project opens in its own tab in the top bar.
“Where is the Launch button in the project bar?” It no longer exists. An empty tab offers the large Launch name button; once an agent is added, you launch from the dock at the bottom of the grid.